Knowledge Injection Method for Real-Time Decision Support

Abstract

Industrial networks are essential components of the infrastructure of industrial and civil facilities, connecting various aspects of a complex social “organism.” Therefore, network damage has serious consequences that can cost money or human lives. Furthermore, in today’s society, the destruction or damage of infrastructure instantly returns populations to the Middle Ages. Therefore, identifying and localizing problems is an essential task of our time.

Modern cities in northern regions use district heating systems to heat their residents and public spaces. Moreover, city residents can purchase heating through pipe networks from combined heat and power plants, nuclear power plants, or waste incineration plants. Therefore, district heating networks are an essential component of modern cities in northern regions. Statistics also indicate that the number of consumers and the length of the networks are increasing. As a result, maintaining district heating network monitoring will be a problem for a very long time. This is why, in this dissertation, the case of district heating network condition monitoring is discussed as a critical infrastructure object.

Technical solutions have been developed to ensure monitoring and emergency response in critical facilities. Expert systems are one example of the data-driven solutions used to monitor the condition of sensor networks. This method uses expert knowledge from the domain, represented as a domain model.

This dissertation discusses a solution based on the analysis of data received from sensors embedded in the district heating network. In order to acquire crucial information to enable real-time decision-making, this dissertation has devised a method for arranging data observations. The approach is tested in a district heating use case, but it can be applied to other situations as well. The approach presented in this study is based on knowledge injection. The developed method separates stable expert knowledge from unstable realtime sensor data. Additionally, the developed method can link and organize searches on structurally different data.